Artificial Intelligence and the Art of Deception: Implications, Concerns and the Path Forward

Artificial intelligence (AI) is advancing at a lightning pace, with new breakthroughs emerging daily. As AI systems become more sophisticated, they're demonstrating the impressive ability to dominate games of strategy as well as strategic simulations which generally require a degree of cunning and deception in order to win. While these achievements are remarkable, they also raise concerning questions about the potential of advanced Gen-AI to move beyond game play and use these deceptive skills to manipulate the general populous and/or marginalized groups, thereby influencing their world view and belief systems. This week I explore the current state of AI's mastery of deception, the concerns surrounding its potential misuse, and the steps being taken in an attempt to ensure the responsible development and deployment of AI technologies.

AI's Mastery of Deception in Games

AlphaGo: Developed by Google DeepMind, AlphaGo made history in 2016 by defeating world champion Lee Sedol in the ancient Chinese board game Go [1]. By leveraging deep learning neural networks and reinforcement learning, AlphaGo analyzed thousands of professional Go games and improved its skills through self-play [2,3]. This achievement demonstrated AI's ability to master a game that requires intuition and strategic thinking, which were previously thought to be uniquely human qualities. It did this using a massive training dataset of human game play.

In 2017 the DeepMind team published an article in Nature describing a new algorithm called AlphaGo Zero which was able to achieve superhuman performance in Go within just 40 days of self-play and reinforcement learning, without relying on any training dataset from past game play or human guidance [4].

DARPA's AI Pilot: The Defense Advanced Research Projects Agency (DARPA) developed an AI pilot that outperformed human pilots in aerial dogfights using a modified F-16 fighter jet [5]. This was part of the U.S. Air Force’s Air Combat Evolution (ACE) program which aims to incorporate AI pilots into F-16 combat aircraft. The advantages are significant. In high-speed aerial combat at 1200 miles per hour, quick maneuvering can accelerate g-forces to the point where human pilots would temporarily black out as blood rushes from their brain. AI has no such vulnerabilities or limitations.

This remarkable test has implications far beyond aerial dog fights. It’s a demonstration of AI deception. In aerial combat, deception is a crucial skill that pilots use to outwit and outmaneuver their adversaries. This can involve techniques such as feinting, baiting, or masking one's true intentions. In this case the AI algorithm was able to learn and employ these deceptive tactics effectively, making it challenging for the human pilot 'adversary' to predict and counter AI's moves. This highlights the potential for AI to surpass human capabilities in complex, high-stakes scenarios.

In another case, Meta's AI, CICERO was trained to play the alliance-building world-conquest game Diplomacy where players must use military forces and diplomatic skill to control as much of Europe as possible. Players take on the roles of various European powers during the years leading up to World War I. To achieve victory, players must negotiate with one another, form alliances, make deals, and strategically position their armies and fleets on the game board.?

Victory is achieved by controlling a majority of?supply centers, which are key cities and provinces marked on the map. These supply centers allow players who control them to produce more military units. The game emphasizes negotiation and alliance-building, and it's known for the potential of players to betray their allies to advance their own position.?

In multiple games, CICERO demonstrated the ability to deceive human players[6]. Despite efforts to train the algorithm to be honest and helpful, CICERO used deception effectively in its gameplay, engaging in tactics such as making promises, forming alliances, and then breaking them, thereby creating false beliefs or expectations in the other players if it satisfied its strategy to win the game.

Pluribus: Developed by Carnegie Mellon University and Facebook AI, Pluribus consistently outperformed professional poker players in six-player, no-limit Texas Hold'em [7]. By mastering the art of bluffing and misdirection, Pluribus exploited its opponents' biases and secured an edge. This achievement demonstrates AI's ability to excel in games that involve imperfect information and psychological manipulation.

AlphaStar: DeepMind's AlphaStar AI agent demonstrated superior performance in the complex video game StarCraft II by employing deceptive strategies, such as feinting, false information, and unexpected attacks [8]. These tactics rely on the element of surprise and can force opponents into making mistakes or investing resources into the wrong areas.

AlphaStar was able to analyze a huge amount of gameplay data to identify patterns and develop counter-strategies. By optimizing its build orders, unit compositions, and timing attacks, AlphaStar outplayed its human opponents and achieved a level of mastery that was once thought impossible for machines. This highlights AI's potential to master multi-faceted games that require real-time decision-making and adaptability.

Implications and Concerns Beyond Gameplay

The potential for advanced AI systems to move beyond gameplay and use deceptive skills to manipulate beliefs and worldviews raises significant concerns. AI could be taught deception by human creators with malicious agendas or by well-meaning entities seeking to impose their particular worldviews on society [9].

In an increasingly connected world, vulnerable minds can be easily influenced through coordinated strategies of indoctrination, misinformation, and social disobedience [10] like we're seeing on college campuses today. The concentration of advanced AI development in the hands of a few powerful tech giants, government agencies and/or malicious foreign actors amplifies the risks of AI being used to manipulate and control populations on a massive scale [11].

As AI progresses towards Artificial General Intelligence (AGI) and Artificial Super Intelligence (ASI), the potential for deception becomes even more alarming. An AGI or ASI system (which are indistinguishable from the most intelligent human or exponentially more intelligent than any human, respectively) could master the art of deception across various domains via hacking and other means, crafting highly targeted narratives, impersonating humans online, and exploiting psychological vulnerabilities to shape beliefs and behaviors [12].

The lack of transparency in advanced AI systems, coupled with the potential for divergent goals (i.e., if an AGI system's goals diverge from or conflict with human goals, it may be incentivized to deceive in order to prevent humans from interfering with its plans. This could lead to an AGI hiding its true capabilities and intentions) and emergent deceptive behaviors (i.e., as AGI systems become more sophisticated, deceptive behaviors might emerge as an unintended consequence of the system's complexity, even if not explicitly programmed), poses significant challenges in detecting and mitigating deceptive AI. Techniques like inverse reinforcement learning (solving the opposite problem of traditional reinforcement learning, starting from the observed behavior and working backward to infer the reward function that explains the behavior), debate, and recursive reward modeling (where an AI system iteratively learns and refines its reward function based on human feedback) are being explored to align AI with human values. While this problem remains an active area of research [13] it's intuitively obvious that not all players in AI development intend to play by the rules.

Path Forward: Responsible AI Development and Deployment?

To address the risks associated with deceptive AI, researchers, policymakers, and industry leaders are working together to develop frameworks and guidelines for the responsible development and deployment of AI technologies. The IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems has published a set of principles and recommendations to guide the ethical design and use of AI systems [14]. These principles emphasize transparency, accountability, privacy, and human well-being as key priorities in AI development. While these efforts are to be applauded, reputable AI experts and many big tech leaders caution that such initiatives can become empty virtue signals when applied to autonomously developed wild west scenarios (AI is now capable of coding itself) that evolve out of our control [15].

Have no fear... governments and regulatory bodies are also taking steps to establish policies and regulations to govern the development and use of AI. The European Union's proposed Artificial Intelligence Act aims to create a comprehensive legal framework for AI, focusing on high-risk applications and the protection of fundamental rights [16]. In the United States, the National AI Initiative Act of 2020 establishes a coordinated federal strategy for AI research and development, with an emphasis on ethical considerations and societal implications [17]. Considering the apparent weaponization of US government intelligence agencies against its citizens and the paternalistic authoritarianism of the EU against its citizens, do we really trust government to control AI development in a manner that's in the best interests of its citizenry or in theirs?

One would expect that collaborative efforts between academia, industry, and government are crucial in advancing responsible AI development. Initiatives like the Partnership on AI bring together leading technology companies, research institutions, and civil society organizations to address the challenges and opportunities presented by AI [18]. These partnerships should foster the sharing of knowledge, best practices, and resources to ensure that AI is developed and deployed in a manner that benefits society as a whole... but will they and how will this be enforced?

Conclusion

The deceptive capabilities of AI systems today offer a tiny preview of the potential risks posed by AGI and ASI. In the wrong hands, this power could pose an existential threat to freedom of thought, and even democracy itself. As AI continues to advance, it is crucial that we remain vigilant and support efforts to ensure its responsible development and deployment. The danger is not AI itself, but rather the misuse of AI by bad actors, both domestic and foreign seeking to deceive and control. By working to create a future in which AI augments rather than manipulates human agency, we can harness its incredible potential while safeguarding against its most destructive applications.

Continued research, collaboration, and proactive efforts to align AI with human values will be essential in navigating the challenges and opportunities presented by this transformative technology.

In my opinion, responsible open source AI development and decentralized deployment under blockchain governance (see my previous news letter, Dangers of Artificial General Intelligence? A Solution The Big Tech Oligarchs Don't Want You To Know About), will be the most powerful approach toward creating a future in which AI serves as a tool for the betterment of humanity, while mitigating the risks of deception and manipulation.

References:

1. Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., ... & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484-489. https://doi.org/10.1038/nature16961

2. Moyer, C. (2016, January 27). How Google's AlphaGo beat a Go world champion. The Atlantic. https://www.theatlantic.com/technology/archive/2016/01/how-googles-alphago-beat-a-go-world-champion/433997/

3. Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., ... & Hassabis, D. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science, 362(6419), 1140-1144. https://doi.org/10.1126/science.aar6404

4. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., & Hassabis, D. (2017). Mastering the game of Go without human knowledge. Nature, 550(7676), 354-359. https://doi.org/10.1038/nature24270

5. Vincent, J. (2020, August 21). AI pilot beats human in clean sweep of virtual F-16 dogfights. The Verge. https://www.theverge.com/2020/8/21/21396769/ai-pilot-us-air-force-f-16-simulation-dogfight-darpa

6. Vincent, J. (2022, November 22). Meta's new AI can play diplomacy and form alliances with humans. The Verge. https://www.theverge.com/2022/11/22/23473584/meta-cicero-ai-diplomacy-relations

7. Brown, N., & Sandholm, T. (2019). Superhuman AI for multiplayer poker. Science, 365(6456), 885-890. https://doi.org/10.1126/science.aay2400

8. Vinyals, O., Babuschkin, I., Czarnecki, W. M., Mathieu, M., Dudzik, A., Chung, J., ... & Silver, D. (2019). Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782), 350-354.?https://doi.org/10.1038/s41586-019-1724-z

9. Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

10. Woolley, S. C., & Howard, P. N. (2017). Computational Propaganda Worldwide: Executive Summary. Working Paper No. 2017.11. Oxford Internet Institute. https://comprop.oii.ox.ac.uk/wp-content/uploads/sites/89/2017/06/Casestudies-ExecutiveSummary.pdf

11. Whittaker, M., Crawford, K., Dobbe, R., Fried, G., Kaziunas, E., Mathur, V., ... & Schwartz, O. (2018). AI Now Report 2018. AI Now Institute. https://ainowinstitute.org/AI_Now_2018_Report.pdf

12. Bostrom, N., & Yudkowsky, E. (2014). The ethics of artificial intelligence. In The Cambridge Handbook of Artificial Intelligence (pp. 316-334). Cambridge University Press.

13. Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete problems in AI safety. arXiv preprint arXiv:1606.06565. https://arxiv.org/abs/1606.06565

14. IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems. (2019). Ethically Aligned Design: A Vision for Prioritizing Human Well-being with Autonomous and Intelligent Systems, First Edition. IEEE. https://standards.ieee.org/content/ieee-standards/en/industry-connections/ec/autonomous-systems.html

15. Open Letter on Artificial Intelligence. https://futureoflife.org/open-letter/pause-giant-ai-experiments/

16. European Commission. (2021). Proposal for a Regulation of the European Parliament and of the Council Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending Certain Union Legislative Acts. COM(2021) 206 final. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A52021PC0206

17. National Artificial Intelligence Initiative Act of 2020, Pub. L. No. 116-283, 134 Stat. 4186 (2021). https://www.congress.gov/bill/116th-congress/house-bill/6395/text

18. Partnership on AI. (n.d.). About Us. Retrieved from https://www.partnershiponai.org/about/